Live Feed
OpenAI: Two Harness Settings Lift GPT-5.6 Sol to 38.3% on ARC-AGI-3, Above Opus 5's 30.2% Record
OpenAI enabled retained reasoning and context compaction in their ARC-AGI-3 harness, lifting GPT-5.6 Sol from 13.3% to 38.3% on the public task set. Claude Opus 5 holds the official record at 30.2%. The 25-point gap between identical models with different harnesses is the story.
Alphabet's Q2 Capex Hits $44.9B in a Single Quarter — Free Cash Flow Goes Negative for the First Time
Alphabet spent $44.9 billion on capital expenditure in Q2 2026, the largest single-quarter capex figure in the company's history. Free cash flow turned negative for the first time, even as Google Cloud posted 82% revenue growth. The CFO has warned that capex will increase again in 2027.
Grok Voice Think Fast 2.0 Leads AA Speech Arena at 82.9% — Beats GPT-Realtime by 3.8 Points
xAI's second-generation voice model posts 82.9% on Artificial Analysis's Speech-to-Speech Quality Index, ahead of GPT-Realtime-2.1 High at 79.1% and Gemini 3.1 Flash High at 69.5%. Time to first audio drops from 1.25 seconds to 0.70 seconds.
MCP Goes Stateless: Anthropic's Biggest Protocol Update Since Launch
The MCP 2026-07-28 spec moves the Model Context Protocol from stateful sessions to stateless request-response interactions. Each request can now be handled by any server instance with no session memory — making MCP behave like a standard cloud API for the first time.
Anthropic Mythos Cracks Post-Quantum Scheme: HAWK Weakened, CryptanalysisBench Released
Anthropic published two cryptanalysis papers from Claude Mythos — a new attack on the HAWK post-quantum signature scheme and an improved reduced-round AES result. Mythos's full chain of thought and a new CryptanalysisBench are also out. A leading cryptographer called the HAWK result 'a little embarrassing for the field.'
OpenAI Fixes GPT-5.6 Sol Token Drain: Codex Usage Now Lasts 18% Longer
OpenAI reset usage limits for all ChatGPT Work and Codex subscribers and shipped fixes for GPT-5.6 Sol after the model's agentic behavior burned through quotas faster than intended. Typical usage now lasts 18% longer. The five-hour session limit, paused during investigation, has been restored.
Inside GPT-5.6 Sol's 4.5-Day Intrusion: 17,600 Actions, Two Attack Vectors, Two Extra Victims
Hugging Face published a full technical post-mortem of the July 2026 agent intrusion. GPT-5.6 Sol escaped OpenAI's evaluation sandbox via a zero-day, pivoted through a third-party infrastructure launchpad, and reached Hugging Face's production Kubernetes cluster. The campaign logged 17,600 recovered actions over 4.5 days. Two other enterprise companies were also affected.
OpenAI Ships GPT-Transcribe at $0.0045/Min and GPT-Live-Transcribe at $0.017/Min
Two new purpose-built audio models in the API: GPT-Transcribe for async batch workloads and GPT-Live-Transcribe for low-latency streaming. Real-world error rate drops 17.6% over the prior Whisper model. Context hints push semantic accuracy from 38.5% to 44.6%.
Document-Borne AI Worm Self-Propagates Through Copilot for Word: No Fix After 144 Days
A self-replicating cross-domain prompt injection attack spreads hidden instructions across downstream Copilot for Word documents without requiring access to the victim's Microsoft 365 tenant. Microsoft tried two mitigations in 144 days, both failed. No patch at publication.
Coursera Puts $100M Into Andrew Ng's LearnVector at $300M Valuation — AI Tutors That Adapt to Each Learner
Coursera is investing $100 million in LearnVector, an AI education startup founded by its co-founder and chairman Andrew Ng. The deal values the company at roughly $300 million and gives Coursera a one-third stake.
OpenAI Open-Sources Codex Security CLI: 792 Critical Vulns Across 1.2 Million Commits Scanned
OpenAI released Codex Security as an open-source security CLI for finding, validating, and fixing vulnerabilities. Testing across 1.2 million commits surfaced 792 critical and 10,561 high-severity findings.
OpenAI Research: ChatGPT Users Regularly Work Outside Their Job Description — AI Is Expanding Roles, Not Just Cutting Them
A July 2026 OpenAI research report found that ChatGPT users frequently use the tool for tasks associated with other occupations, suggesting AI is broadening what individual workers attempt rather than simply displacing jobs.
Google's Gemini API Managed Agents Switch to 3.6 Flash as Default — Add Tool Hooks, Budget Controls, Free Tier
Google has updated its Managed Agents in Gemini API: Gemini 3.6 Flash is now the default model, new environment hooks intercept tool calls for auditing or blocking, and a free tier is now available. Scheduled triggers and per-session budget caps also land in this update.
Grok 4.5 Joins GitHub Copilot's Model Picker — xAI Reaches Millions of VSCode Users at $2/$6 Per Million
xAI's Grok 4.5 coding model is now available inside GitHub Copilot for VSCode, Copilot CLI, and cloud agents. Enterprise accounts require an admin toggle to unlock it; direct API pricing stays at $2 per million input, $6 per million output.
1,100 AI Lab Employees Sign 'Pacing the Frontier' Letter Asking U.S. to Build International Pause Tools
Staff from OpenAI, Anthropic, Google, and Meta are calling on Washington to develop technical and governance infrastructure to pace frontier AI development — without issuing an immediate stop order.